基于元学习的时变非线性负荷预测组合算法

单一的预测算法或多或少存在着归纳偏置,由此导致了系统偏差的普遍性。提出了一种基于元学习的时变非线性组合预测算法,该算法在进行组合预测时将序列的特征属性和基预测器预测的结果形成元知识,作为元预测器的输入,从而发现并且纠正基预测器的系统偏差。在元预测器中,通过门控网络确定各基预测器的权重,保证了权重的时变性和非负性。将该算法应用于电力负荷超短期预测,预测结果表明,该算法的预测精度高于单一预测算法和常用的线性和非线性组合算法。

Time-varying nonlinear power load combined forecasting algorithm based on meta-learning

Inductive bias exists in single prediction algorithm more or less, which results in system bias usually. A new time-varying nonlinear combined forecasting algorithm is presented. Meta knowledge formed by the results of base predictors and feature attributes of series is used as inputs of meta predictor when combined forecasting is applied. System bias can be founded and rectified. The weights of base predictors are calculated using gating network in metal predictor. Weights of base predictors are time-varying and non negative. The new algorithm is applied in very short-term load forecasting. Results show that the proposed method improves forecasting precision comparing with single prediction algorithm and normal combined forecasting algorithm.